[molpro-user] GPGPU calculation for Molpro?

[Gerald Knizia]

On Monday 29 June 2009 12:13, Gang wrote:
> It is about the GPGPU calculation.  People claim that this graphic
> processor associated calculation could increase the speed of some
> programs by a factor of 5 - 100, which is quite impressive. And PGI just
> released a new compiler base on this new technology.   

There certainly is substantial potential in GPGPU application. However, the 
area with the greatest benefits are clearly single precision calculations 
which involve very large matrices (or other massive amounts of data with few 
interdependencies manipulated in exactly the same way for all the data).

Some hardware developers I spoke to assured me that it will stay like this for 
quite some time. Mainly because some double precision arithmetic units 
require 4 times as much chip space as single precision units, and for the 
main application area of the cards---computer graphics--single precision is 
perfectly sufficient.

> The GPGPU  calculation seems quite good at matrix operation and
> image processing. 
> Which kind of operations are molpro doing?  Will GPGPU calculation 
> increase the speed of molpro quite a lot?

Molpro is doing lots of matrix operations, but most of them involve matrices 
which are too small for an efficient use of GPUs. A straightforward exchange 
of a normal BLAS with a GPU based BLAS, for example, has been tried, but 
currently does not offer substantial increases in computation time relative 
to a Multi-Core BLAS or pure parallel execution on the CPU.

This may change in the future when the software can be adapted to the new 
hardware developments. But this is a very complex and complicated endeavor 
and will require some time.
-- 
Gerald Knizia